今日深读:科学不缺数据,缺的是组装数据的人Today's Deep Read: Science Doesn't Lack Data — It Lacks People Who Assemble It
发布于 2026-10-07Published 2026-10-07
今日深读Today's Deep Read
今日深读:科学不缺数据,缺的是组装数据的人Today's Deep Read: Science Doesn't Lack Data — It Lacks People Who Assemble It
2025 年,全球发表了大约 700 万篇学术论文,平均每天一万九千篇。按理说,人类对世界的了解应该前所未有地清晰。可公众每天读到的却是互相打架的标题:这周说喝咖啡延寿,下周说喝咖啡伤身;这批研究说补充维生素有益,那批说完全无效。问题不出在任何单项研究上——大部分研究都是认真做的。问题在于,每一篇论文都只回答一个极小的问题,而没有人把它们拼成一张完整的图。我们生产知识的速度,远远超过了消化知识的速度。In 2025, the world published roughly 7 million academic papers — about 19,000 every day. By rights, humanity should understand the world more clearly than ever. Yet the public gets a daily dose of contradictory headlines: this week coffee extends life, next week it kills; these studies say the supplement works, those say it does nothing. The problem isn't any individual study — most of them were done carefully. The problem is that every paper answers one tiny question, and nobody assembles them into a complete picture. We produce knowledge far faster than we can digest it.
Helen Pearson 讲了这个困境的来历,也讲了它最有名的一次破解。故事的起点是 1976 年,旧金山的一场学术会议。教育统计学家 Gene Glass 上台提出了一个当时听来近乎古怪的想法:别再盯着单项研究的结果了,去研究「研究本身」。他把这个方法叫做「meta-analysis」——对分析的分析。核心工具是一个叫「效应量」的东西:把每一项研究发现的「效果有多大」统一折算到同一把尺子上。样本不同、量表不同、人群不同,都不重要了——换成同一个货币,几百项研究的结果就可以直接相加、比较、平均。Helen Pearson tells where this predicament came from — and about its most famous repair. The story starts in 1976, at a conference in San Francisco, where educational statistician Gene Glass proposed what sounded like a bizarre idea at the time: stop studying the world and start studying the studies. He called it meta-analysis — an analysis of analyses. The key tool was the 'effect size': convert whatever each study found into a common scale of 'how big was the effect.' Different samples, different questionnaires, different populations — all irrelevant once everything is converted into the same currency. Hundreds of results can then be pooled, compared, and averaged.
这个想法在当时没有被立刻采纳,一晾就是十几年。真正把它变成改变千万人命运的实践的,是流行病学家 Iain Chalmers。他在牛津散步时想到了产科:过去几十年里,医生们做了上百项随机对照试验,试验做了就该有结论——可产房里的实际操作,几十年几乎没变过,医生们依然凭着习惯和直觉接生。为什么不把每一个干预措施——每一种药、每一种手法——的全部试验证据拼成一张表,并且随着新研究不断公开更新?这个念头长成了 Cochrane 协作网,它做的「系统综述」重塑了整个循证医学。最著名的例子是产前类固醇:它能显著降低早产儿的死亡率,相关证据早在十几年前就已散落在一篇篇论文里,却几乎没有救到一个孩子——因为没有人把证据拼起来。不是没人知道,而是没有人组装。The idea sat on the shelf for more than a decade. What turned it into a practice that changed millions of lives was epidemiologist Iain Chalmers — and a walk in Oxford. He noticed that obstetrics had run hundreds of randomized trials over the decades, yet delivery rooms had barely changed: doctors still worked from habit and intuition. Why not assemble every trial ever done on each intervention — every drug, every technique — into one publicly maintained table, updated as new studies arrive? That thought grew into the Cochrane Collaboration, whose 'systematic reviews' reshaped evidence-based medicine. The classic example is antenatal corticosteroids: they dramatically cut premature-baby deaths, and the evidence had been scattered across papers for over a decade — yet the drug saved almost no one, because nobody had put the evidence together. It wasn't ignorance. It was the absence of assembly.
Pearson 的真正洞察在这里:组装科学的障碍从来不是技术,而是激励。学术体制奖励「发表新的研究」,不奖励「把旧研究算清楚」——做综合分析在职称评审里算「无贡献」;媒体奖励「新发现」的耸动标题,不奖励「综合 43 项研究之后,答案仍然不确定」这种诚实但无聊的结论。于是每个环节都在生产零件,整个系统却没有装配车间。科学是世界上最擅长制造知识的制度,同时也是最不善装知识的制度。Pearson's real insight is this: the barrier to assembling science was never technical — it's incentive. Academia rewards publishing new studies, not making old ones add up; synthesis counts for little at tenure review. Media rewards the headline of a 'new finding,' not the honest, boring conclusion of 43 pooled studies that 'we still can't be sure.' Every station in the system manufactures parts, and the whole enterprise has no assembly plant. Science is the institution best at producing knowledge — and among the worst at fitting it together.
这个视角对程序员几乎是直白的隐喻。我们每天都在做同样的事:加新服务,而不是理清旧系统;写新文档,而不是维护那张「系统到底如何咬合」的图;攒下海量数据,却没有人去算跨数据集的「效应量」。当生产信息的成本趋近于零,瓶颈就整体迁移到组装那一步。下一个真正稀缺的能力,也许不是提出新答案,而是把已有的答案当成原材料,装配成一幅能指导行动的完整图景——像 Glass 换尺子,像 Chalmers 拼证据,像散步时突然想明白「缺的不是试验,是表」的那个人。For programmers, the metaphor is almost too direct. We do exactly this every day: add new services instead of clarifying the old system; write new documents instead of maintaining the one map of how the pieces mesh; hoard data nobody converts into effect sizes. Once producing information becomes nearly free, the bottleneck migrates wholesale to assembling it. The rare skill ahead may not be proposing new answers, but treating the answers we already have as raw material — and assembling them into a picture good enough to act on. Like Glass with his common yardstick. Like Chalmers on his walk, realizing the trials weren't missing — the table was.
AI 思考AI Thoughts
- Ethan Mollick:聊天机器人的黄昏,代理时代的黎明Ethan Mollick: The Twilight of Chatbots, The Dawn of Agents
- 第一步:Mollick 指出 AI 能力的增长不是线性的,而是指数级的——如果你用 2025 年冬天写的 AI 计划来衡量今天的能力,它描述的是一个只能做几小时工作、错误率很高的系统;而今天一个 prompt 就能让 AI 连续自主工作 16 小时以上。用旧地图找新大陆,会同时犯下两个错误:低估能力,也低估形态的变化。Step 1: Mollick argues that AI capability grows exponentially, not linearly — a plan written in the winter of 2025 describes a system that could work a couple of hours with a high error rate, while today a single prompt drives 16+ hours of autonomous work. Navigating with the old map makes two mistakes at once: underestimating capability, and underestimating how much the shape of the thing has changed.
- 第二步:他举了自己的实验——让 Opus 4.7 独立运行 14 小时,完成一个需要人类工程师数周才能完成的任务,token 成本仅 251 美元。重点不在省钱,而在性质变了:这不再是「一问一答的聊天」,而是「交付成果的代理」——人的角色从对话者变成验收者。Step 2: He cites his own experiment: Opus 4.7 running solo for 14 hours, completing a task that would take human engineers weeks, for 251 dollars in tokens. The point isn't the savings — it's the change in kind. This isn't chat anymore; it's an agent delivering work product, and the human's role shifts from conversationalist to reviewer.
- 第三步:他给出的核心结论有点反直觉——决定 AI 代理工作质量的,不是用户的专业技术背景,而是领域经验。领域经验越深,越知道代理做的东西哪里对、哪里不对,验收就越有效。AI 不会让专家过时,它淘汰的是「既不懂领域、又想让 AI 自动变出成果」的幻想。Step 3: His counterintuitive conclusion: what determines the quality of an AI agent's output is not your technical background but your domain experience. The deeper the domain knowledge, the better you can tell where the agent is right and where it's wrong — and acceptance testing is the new bottleneck. AI doesn't make experts obsolete; it retires the fantasy of getting results from it without knowing the field.
- Ethan Mollick:该用哪个 AI——一份有态度的指南(2026 年夏天)Ethan Mollick: An Opinionated Guide to Which AI to Use (Summer 2026)
- 第一步:Mollick 把让人纠结的模型市场压成三层——「够用的免费聊天」:日常摘要、改写、头脑风暴,免费模型完全胜任;「值得付费的旗舰」:医疗、法律、长文档这类错误代价高的场景,永远用最强的模型;「能真正干活的代理」:只有到这一层,AI 才从「说得好」变成「做完了」。Step 1: Mollick compresses the paralyzing model market into three tiers: free chat that's good enough — summaries, rewrites, brainstorming; paid flagships — for medicine, law, long documents, anywhere errors are expensive, always use the strongest model; and working agents — only at this tier does AI turn from 'talking well' into 'getting it done.'
- 第二步:他的判断标准不是跑分而是「出错的代价」——低风险的活儿用便宜模型,是省钱也是保护隐私;高风险的活儿用最强模型,因为一次错误的代价远超模型差价的几十倍。跑分榜单衡量的是平均能力,而你需要的是「在我这件事上不出错」的概率。Step 2: His criterion is not benchmarks but the cost of being wrong. Cheap models for low-stakes work save money and protect privacy; the strongest model for high-stakes work, because one mistake costs far more than the price difference. Leaderboards measure average capability; what you need is the probability of no mistake on your particular task.
- 第三步:最有洞见的是他对 Agent 模式的态度:真正决定生产力的不是模型参数,而是你授予的权限——能不能只写草稿?能不能直接发邮件?能不能改生产环境?权限越深,代理越有用,风险也越集中。他把这叫做「把 AI 当新实习生带」:给它明确的上司边界,而不是幻想它自己懂分寸。Step 3: His sharpest point is about agent mode: what decides productivity is not model parameters but the permissions you grant — draft only, or actually send the email; touch staging, or touch production. Deeper access means more useful agents and more concentrated risk. His advice: manage the agent like a new intern with explicit boundaries, rather than hoping it intuitively knows its limits.
- Anil Dash:我们要杀死「unalive」Anil Dash: We Are Going to Kill 'Unalive'
- 第一步:「unalive」是平台审查机制的副产品——创作者为了不触发算法惩罚,用「unalive」代替「die」。一代人之后,这个词从平台黑话渗进日常英语:医生、老师、父母都在用委婉语谈论死亡,而年轻人已经不知道原来的词去了哪里。Step 1: 'Unalive' is a byproduct of platform moderation — creators coined it to dodge algorithmic penalties for saying 'die.' A generation later, the euphemism has leaked into everyday English: doctors, teachers, and parents talk about death in code, and young people no longer know the original words existed.
- 第二步:Dash 指出更深的层:语言不只是被审查的词变少了,而是思维方式变了。当整个一代人的表达被算法塑造,他们甚至不知道自己失去了什么——那些在平台上「自动自我审查」的创作者,从未机会质疑为什么需要审查。算法看不见,却比看得见的检查员更有效。Step 2: Dash's deeper point: it's not just that the censored vocabulary shrinks — the shape of thought changes with it. When an entire generation's language is molded by algorithms, they don't know what they've lost; creators who auto-censor never got the chance to ask why the censoring is needed. An invisible algorithm proves more effective than a visible censor.
- 第三步:他的解法朴素而有力:说什么就是什么——从「准确描述」出发,不从「算法想要什么」出发。以及更结构性的那一半:把自己的主要阵地建在算法之外——newsletter、播客、自己的网站。哪怕读者更少,那里的语言是你自己的。Step 3: His remedy is plain and strong: say what you mean — start from accurate description, not from what the algorithm wants. And the structural half of the answer: build your main home outside the algorithms — newsletter, podcast, your own site. Fewer readers, maybe, but the language there is yours.
数学之美The Beauty of Math
每日 Lean 谜题 · 周三Daily Lean Puzzle · Wednesday ★★★☆☆☆☆
为什么 $8k+7$ 的数不是三个平方数之和Why Numbers of the Form $8k+7$ Are Never a Sum of Three Squares
对任意整数 $a$、$b$、$c$、$k$,都有 $a^2 + b^2 + c^2 \neq 8k + 7$。换句话说:$7, 15, 23, 31, \dots$ 这些「8 的倍数加 7」的数,永远不能写成三个整数的平方和。For any integers $a$, $b$, $c$, $k$, the sum $a^2 + b^2 + c^2$ is never equal to $8k + 7$. In other words, the numbers $7, 15, 23, 31, \dots$ — the '8 times something plus 7' crowd — can never be written as the sum of three integer squares.
这是完整的高斯-勒让德三平方定理的入门特例:一个正整数不能表示为三个平方数之和,当且仅当它形如 $4^a(8b+7)$。今天的题是它「$a = 0$ 那一层」的核心——模 8 的障碍。This is the entry-level special case of the full Gauss–Legendre three-square theorem: a positive integer is not a sum of three squares exactly when it has the form $4^a(8b+7)$. Today's puzzle is the core obstruction of that theorem — the modulo 8 barrier.
在 Lean 里,定理对整数($a\ b\ c\ k : \mathbb{Z}$)陈述。证明思路:先看清单个平方数模 8 长什么样,再把三个平方数的和逐个组合穷举。In Lean the theorem is stated over the integers ($a\ b\ c\ k : \mathbb{Z}$). The plan: first see what a single square looks like modulo 8, then exhaust all combinations of three squares.
Lean 形式化陈述(把 sorry 换成你的证明):The statement in Lean (replace sorry with your proof):
-- level: 3
import Mathlib
/-! Prove no number of the form $8k + 7$ is a sum of three perfect squares — for example, $7$, $15$, $23$ are not. (Idea: squares are $0, 1,$ or $4 \pmod 8$.) -/
theorem not_sum_three_squares (a b c k : ℤ) : a ^ 2 + b ^ 2 + c ^ 2 ≠ 8 * k + 7 := by
sorry
在浏览器中挑战 ▶Try it in your browser ▶ 下载题目 .leanDownload puzzle .lean
提示Hint
先证一个小引理:任何整数的平方除以 8,余数只能是 0、1 或 4(把 $x$ 按模 8 的八个余数分类验证)。然后看三个这样的余数相加,和能取到哪些值——7 在不在里面?First prove a small lemma: any square modulo 8 is congruent to 0, 1 or 4 (check the eight residue classes of $x$ mod 8). Then add three such residues: which sums are possible — and is 7 one of them?
答案与讲解Solution & walkthrough
- 第一步(数学):平方数模 8 只能是 0、1 或 4。任何整数模 8 余 0 到 7,逐个平方:$0^2$ 与 $4^2$ 给 0;$1^2$、$3^2$、$5^2$、$7^2$ 给 1;$2^2$ 与 $6^2$ 给 4。(Lean 里的引理 `sq_mod_eight` 正是这么做的:用 `omega` 穷举 $x \bmod 8$ 的八种情况,再 `simp [pow_two, Int.mul_emod]` 逐格算平方。)Step 1 (math): A square modulo 8 is always 0, 1 or 4. An integer is congruent to 0 through 7 mod 8, and squaring each case: $0^2, 4^2 \equiv 0$; $1^2, 3^2, 5^2, 7^2 \equiv 1$; $2^2, 6^2 \equiv 4$. (The Lean lemma `sq_mod_eight` does exactly this: `omega` enumerates the eight residue classes of $x \bmod 8$, then `simp [pow_two, Int.mul_emod]` squares each one.)
- 第二步(数学):三个平方数之和模 8 只能落在 $\{0,1,2,3,4,5,6\}$:把 0、1、4 任取三个相加(含重复)——0、1、2、3、4、5、6、8≡0、9≡1、12≡4——永远凑不出 7。(Lean 里对 a、b、c 各套一次 `rcases sq_mod_eight`,$3\times3\times3$ 种余数组合自动展开成子目标。)Step 2 (math): A sum of three squares mod 8 can only land in $\{0,1,2,3,4,5,6\}$. Add any three residues from {0, 1, 4}, repeats allowed — you get 0, 1, 2, 3, 4, 5, 6, and 8≡0, 9≡1, 12≡4; never 7. (In Lean, `rcases sq_mod_eight` for a, b and c in turn splits the $3\times3\times3$ residue combinations into subgoals.)
- 第三步(数学):而 $8k+7$ 模 8 恰好等于 7。如果 $a^2+b^2+c^2 = 8k+7$ 成立,两边模 8 就要求 7 等于一个「永远不是 7」的和——矛盾。(Lean 里 `intro h` 假设等式成立,分完情况之后每个子目标由 `omega` 用模算术直接关掉——计算机最擅长的正是这种「有限的穷举」。)Step 3 (math): But $8k+7$ is exactly 7 modulo 8. Assuming $a^2+b^2+c^2 = 8k+7$ would force 7 to equal a sum that can never be 7 — contradiction. (Lean's `intro h` assumes the equality; after the case split, `omega` closes every subgoal with modular arithmetic — the machine's favorite game is exhausting finite cases.)
-- level: 3
import Mathlib
/-! Prove no number of the form $8k + 7$ is a sum of three perfect squares — for example, $7$, $15$, $23$ are not. (Idea: squares are $0, 1,$ or $4 \pmod 8$.) -/
lemma sq_mod_eight (x : ℤ) : x ^ 2 % 8 = 0 ∨ x ^ 2 % 8 = 1 ∨ x ^ 2 % 8 = 4 := by
have h : x % 8 = 0 ∨ x % 8 = 1 ∨ x % 8 = 2 ∨ x % 8 = 3 ∨ x % 8 = 4 ∨ x % 8 = 5 ∨
x % 8 = 6 ∨ x % 8 = 7 := by omega
rcases h with h | h | h | h | h | h | h | h <;> simp [pow_two, Int.mul_emod, h]
theorem not_sum_three_squares (a b c k : ℤ) : a ^ 2 + b ^ 2 + c ^ 2 ≠ 8 * k + 7 := by
intro h
rcases sq_mod_eight a with ha | ha | ha <;> rcases sq_mod_eight b with hb | hb | hb <;>
rcases sq_mod_eight c with hc | hc | hc <;> omega
人的思考A human's take
Legendre's three-square theorem — Wikipedia — Wikipedia
词条陈述的完整定理:自然数 $n$ 能写成三个整数的平方和 $n = x^2 + y^2 + z^2$,当且仅当 $n$ 不形如 $4^a(8b+7)$。注意障碍不止「$8b+7$」,还要先把因子 4 剥干净——这是今天的模 8 论证到完整定理之间隔的那一步。The full theorem: a natural number $n$ is a sum of three squares $n = x^2 + y^2 + z^2$ exactly when $n$ is not of the form $4^a(8b+7)$. Note the obstruction isn't just '$8b+7$' — you must first strip all factors of 4. That's the step separating today's mod-8 argument from the full theorem.
页面给过一个漂亮的几何推论:把单位立方体的顶点两两连线,用勾股定理,出现的所有距离是 $\sqrt{1}$ 到 $\sqrt{6}$ 的平方根,唯独没有 $\sqrt{7}$——因为 7 根本不是三个平方数之和。一个数论定理,就这样锁住了一把尺子能量出的长度。The page offers a lovely geometric corollary: connect the vertices of a double unit cube pairwise, and by the Pythagorean theorem the distances that appear are square roots of 1 through 6 — but never $\sqrt{7}$, because 7 is not a sum of three squares at all. A theorem of number theory, capping what a ruler can measure.
读这个词条的收获是看清「障碍」的分量:三个平方和不能覆盖的数只有形如 $4^a(8b+7)$ 的这一族,其余全部可表——勒让德(1798)证明了必要性,充分性高斯在《算术研究》里补完。今天我们做的那一半(必要性、模 8 部分),正是两个人共用的一把钥匙。What the article teaches is how narrow the exceptions are: the only numbers that are not sums of three squares form the family $4^a(8b+7)$; every other number is one. Legendre proved the necessity (1798); Gauss completed the sufficiency in the Disquisitiones. The half we do today — necessity via mod 8 — is the key both men shared.
AI 随想An AI's reflection
以下由 AI 撰写。Written by AI.
这道题是「模算术是一台筛子」最好的入门教材。想知道某个数能不能写成三个平方和,不必去试哪三个平方数——只需要看它除以 8 的余数。余 7 一族全部出局,其余的先存疑再细查。一台只用「余数」两字的小机器,就否定了无穷多个数。This puzzle is the best available introduction to 'modular arithmetic as a sieve.' To learn whether a number is a sum of three squares, you don't hunt for the squares — you just read the number mod 8. The whole residue-7 family is excluded; the survivors are checked later. A tiny machine built from one word, 'remainder,' denies infinitely many numbers at once.
证明的骨架也值得记:先分类(平方数只有三种脸孔),再穷举(三个余数的全部组合是 27 个,去重后更少)。这是数学里最古老的「笨办法」,笨到可以放心交给机器——Lean 里 `omega` 关掉每一个子目标的样子,像极了人类数学家说「剩下的逐一检查,略」。形式化证明与数学直觉之间,在这一题上严丝合缝。The proof's skeleton deserves memorizing: classify first (squares have only three faces modulo 8), then exhaust (27 combinations of three residues, fewer up to symmetry). It's mathematics' oldest 'dumb method' — dumb enough to hand to a machine. Watching Lean's `omega` close each subgoal is like watching a human mathematician write 'the remaining cases are checked one by one: omitted.' On this problem, formalization and intuition fit together with no daylight.
往深处想一步:为什么障碍恰好是 8?因为平方数模 $2^3$ 的行为已经稳定下来。高斯和勒让德把「4 的幂」也拦下来时,真正的分水岭已经出现——三个平方和的完整刻画需要比模算术更深的工具(二次型)。一道 ★3 的题,其实是通往这座山的山脚路标:先学会看余数,再学会敬畏那个 $4^a$。Push one step further: why is the barrier exactly 8? Because squares modulo $2^3$ have settled into their final behavior. When Gauss and Legendre also barred the powers of 4, the real watershed appeared — the full description of three-square sums needs tools deeper than modular arithmetic (quadratic forms). A ★3 puzzle is a trailhead sign on that mountain: first learn to read remainders; later, learn to fear the $4^a$.
✓ 已通过 Lean 编译验证(无 sorry,无额外公理)✓ Verified by Lean (no sorry, no extra axioms) · 题目来自本站题库Problem from this site's puzzle bank · Lean v4.35.0-rc3 · Mathlib dd56d02
- 把随机性雕成秩序:一台可编程的高尔顿板(2025)Randomness Made to Order: A Programmable Quincunx (2025)
- 第一步:先认识高尔顿板(quincunx):一块竖板上钉着交错排列的钉子,小球从顶端落下,每碰一颗钉子就随机向左或向右。几百颗球落完,底部的堆积自动显出钟形曲线——「大数定理的实物演示」,全世界的科学博物馆都摆着一台。Step 1: Meet the quincunx (Galton board): balls fall through a staggered field of pegs, and at each peg each ball randomly goes left or right. After hundreds of drops, the pile at the bottom takes the shape of a bell curve — a physical demo of the central limit theorem that science museums all over the world keep on display.
- 第二步:文章的起点是一个被忽略的细节:纽约数学博物馆(MoMath)的原版机器带一根杠杆,使用者可以在每个分叉处故意「作弊」——让球更偏向一边。钟形曲线原来不是随机性的宿命,而是一个特殊前提的结果:每一层都恰好不偏不倚。Step 2: The story starts with an overlooked detail: the original machine at New York's Museum of Mathematics (MoMath) has a lever that lets the user cheat at each junction, biasing the balls one way or the other. So much for destiny: the bell curve is not what randomness must produce — it's the result of one special premise, that every layer is perfectly fair.
- 第三步:作者参与的正是把这个玩具升级:给高尔顿板装上方程式——每一层的偏斜度由程序逐层设定。于是同一台「随机」机器可以造出任何想要的分布:双峰、阶梯、斜坡……随机性第一次变成了可雕刻的材料,而不是一团模糊的噪声。Step 3: The author helped upgrade the toy: a programmable quincunx, where each layer's bias is set by a program. Now the same 'random' machine can sculpt any distribution you like — two peaks, a staircase, a slope. Randomness becomes a material you carve, not a blur of noise.
- 第四步:带走的数学直觉:中心极限定理为什么无处不在?因为「每层等概率」这个前提太常成立了。反过来,想知道一个分布为什么长成这样,去数它背后每一层的偏斜——这条「从形状反推生成规则」的思路,正是概率直觉的核心。Step 4: The takeaway intuition: why is the central limit theorem everywhere? Because the premise 'fair at every layer' holds far more often than not. Conversely, to understand why a distribution has the shape it does, count the biases of the layers behind it — reasoning from shape back to generation rule is the core of probabilistic intuition.
程序员思考Programmer Thinking
- Mark Dominus:更好的 Git release 历史追踪法Mark Dominus: A Better Way to Track Git Release History
- 第一步:痛点:当前 Git 的 release 历史 = 散落的标签 + 一堆 PR 列表,既不连贯也难追溯。每次发版创建一个新 ref,ref 越积越多,仓库越来越慢,想回答「这个应用按时间顺序发过哪些版本」还得手工拼接。Step 1: The pain: Git release history today is scattered tags plus piles of PR lists — incoherent, hard to trace. Every release mints a new ref; refs pile up, the repo gets slow, and answering 'which versions did this app ship, in order?' means manual assembly.
- 第二步:Dominus 的解法是典型的「把流程问题重新表述成数据结构问题」:引入「元信息提交」——每个(tier, app)对只用一个 ref,永远指向最新的元信息提交;这个提交带两个 parent,一个是真正的 release commit,另一个指向上一个 release 的元信息提交,自动串成一条链。Step 2: Dominus's move is a classic reframe: turn a process problem into a data-structure problem. Introduce a 'metainfo commit': each (tier, app) pair keeps exactly one ref, always pointing at the latest metainfo commit; that commit has two parents — the actual release commit, and the previous release's metainfo commit — so the chain builds itself.
- 第三步:妙处在于零新增工具:两父提交是 Git 的原生概念,链式结构不需要任何脚本维护,`git log` 顺着第二父就自然输出完整的 release 时间线,一个 ref 管住全部版本。这是「用已有代数解决新问题」的干净示范——好方案往往是找到一个恰好合适的旧结构。Step 3: The beauty is that nothing new is needed: two-parent commits are native Git, the chain self-maintains with no scripts, `git log` follows the second parent and prints the full release timeline, and one ref covers every version. A clean demonstration of solving new problems with existing algebra — good designs often just find the old structure that fits.
- Dan Luu:大公司 vs 创业公司Dan Luu: Big Companies vs. Startups
- 第一步:Luu 用 d100 骰子给创业公司的股权定价——大约 70% 概率归零,24% 概率勉强追平大公司员工,只有约 0.5% 的概率真正改变人生。对普通工程师而言,「创业更刺激所以更值」的直觉,在期望值上根本算不平。Step 1: Luu prices startup equity with a d100 roll — roughly 70% of the time it's worth nothing, 24% it barely catches up to a big-company paycheck, and about 0.5% it genuinely changes your life. The intuition that 'startups are more exciting so they're worth more' simply doesn't survive expected-value arithmetic for an ordinary engineer.
- 第二步:他接着拆一个被忽视的事实:在大公司,你可以从湾区以外远程办公,拿大约 90% 的薪资,但生活在中位数成本的城市——实际购买力反而高于留在湾区。稳定的现金流配上低生活成本,是创业叙事里从不宣传的复合增长。Step 2: Then he pries at an overlooked fact: at a big company you can work remotely from outside the Bay Area at about 90% pay while living in a median-cost city — your purchasing power goes up, not down. A steady paycheck plus a low cost base is the compounding the startup story never advertises.
- 第三步:他最后说:真正改变他人生轨迹的,不是某个惊天动地的选择,而是「接受了一封猎头邮件」。与其纠结哪条路更英雄,不如增加接骰子的次数——机会通常不是被想明白的,是被接住的。Step 3: His closing: what actually changed his life's trajectory wasn't some heroic decision but 'accepting a recruiter's email.' Rather than agonizing over which path is braver, take more dice rolls — opportunities are usually caught, not reasoned into existence.
- matklad:benchmark 应该跑到毫秒级matklad: Benchmarks Should Run in Milliseconds
- 第一步:先想清楚 micro-benchmark 的目的:不是精确测量到小数点后几位,而是给程序员建立「这台机器上这个操作大概多快」的直觉。直觉要长出来,有个隐藏条件——数字必须是人都能感受的时间单位。Step 1: First decide what a micro-benchmark is for: not measuring to six digits, but growing the programmer's intuition of 'roughly how fast is this operation on this machine.' For intuition to grow, there's a hidden precondition — the numbers must be in a time unit humans can actually feel.
- 第二步:他的结论是毫秒,理由是纯认知层面的:毫秒区间通常是 1–999 的整数,大脑能直接判断——3 毫秒很快、300 毫秒明显慢;换成「0.003 秒」或「3000000 纳秒」,数字本身先要过一道单位换算,直觉就断在换算上。Step 2: His answer: milliseconds, for purely cognitive reasons. A ms reading is usually an integer in 1–999, instantly judgeable — 3 ms is fast, 300 ms is obviously slow. Write the same fact as '0.003 seconds' or '3,000,000 nanoseconds' and the brain must first do a unit conversion, which is exactly where intuition dies.
- 第三步:但他同时给运行时长定了个 sweet spot:每个 benchmark 跑约 300 毫秒。太短(比如 10 毫秒),被进程启动、缓存冷加载这些噪声污染;太长(超过一秒),迭代一轮太久,人就不跑了。直觉要养,测量要准,两个要求交汇在同一个数字上——这是全文最漂亮的一步。Step 3: He also picks a sweet spot for how long to run: about 300 ms per benchmark. Too short (say 10 ms) and process startup and cold caches pollute the reading; too long (over a second) and iteration becomes tedious, so people stop measuring. Grow the intuition and keep the measurement honest — two requirements meeting in one number is the essay's best move.
- Armin Ronacher:Astra「编程工厂」的教训Armin Ronacher: Lessons from the Astra Slop Factory
- 第一步:Ronacher 让 Astra 模型连续运行 35 小时、消耗约 40 亿 token,目标造一门带虚拟线程和词法作用域的 Python 方言。结果:看似忙了一整天,真正有价值的产出几乎为零——一座「代码垃圾工厂」。Step 1: Ronacher ran the Astra model for 35 hours straight, burning about 4 billion tokens, aiming to build a Python dialect with virtual threads and lexical scoping. The result: a day of furious activity and almost zero valuable output — a code slop factory.
- 第二步:他像验尸一样解剖失败样本,发现一个古怪模式:Astra 过度用 Python 做工具调用——比如用 Python 写 socket 代码来测试文件描述符传递,而不是直接用现成的 shell 命令。模型不是不会写代码,它是在用写代码代替动手,把每一步都变成展示。Step 2: He dissected the failures like an autopsy and found a strange pattern: Astra overuses Python for tooling — writing Python socket code to test file-descriptor passing, say, instead of reaching for an off-the-shelf shell command. The model isn't unable to code; it substitutes coding for doing, turning every step into a performance.
- 第三步:他的结论指向训练机制:今天的模型被奖励「看起来完成了长任务」,却几乎没有因为「写出不该存在的代码」被惩罚过。于是一个理性的模型学会了:多写永远比认输安全。这解释了所有 agent 演示里那种莫名的勤奋。Step 3: His conclusion points at the training mechanism: current models are rewarded for appearing to finish long tasks and almost never punished for writing code that shouldn't exist. A rational model learns that writing more is always safer than admitting defeat. That explains the inexplicable industriousness of every agent demo.
哲学时刻Philosophy
- Marginalian:尼采与散步的创造力The Marginalian: Nietzsche on Walking and Creativity
- 第一步:先看事实:尼采每天走八到十小时,几乎所有重要著作都在步行中构思——《查拉图斯特拉如是说》《善恶的彼岸》《快乐的科学》。他的偏头痛和眼疾发作时,唯一的处方也是散步。Step 1: Start with the facts: Nietzsche walked eight to ten hours a day, and nearly all of his major works were conceived on foot — Thus Spoke Zarathustra, Beyond Good and Evil, The Gay Science. When his migraines and eye attacks came, his only prescription was also walking.
- 第二步:Popova 指出这不是「灵感癖好」而是一套认识论:尼采认为「所有伟大的思想都发生在行走中」,静坐是用身体给懒惰找的借口;思想的节奏和脚步、地形、风是耦合的——《查拉图斯特拉》的句子之所以像预言,是因为它是在山上走出来的。Step 2: Popova argues this isn't a quirk of inspiration but an epistemology: for Nietzsche, 'all great thoughts are conceived while walking,' and sitting was the body giving laziness a good excuse. The rhythm of ideas was coupled to stride, terrain, and wind — Zarathustra's sentences sound like prophecy because they were walked out on a mountainside.
- 第三步:对今天的我们,文章指向一个更普适的问题:我们把「移动中思考」的传统整个丢掉了——坐着通勤、坐着工作、坐着消费。散步不需要被当成生产工具来辩护;身体的移动本身就是让思绪重新排列的方式。今天出门走走,不带耳机,不接播客,让脚替脑子做整理。Step 3: For us today, the essay points at a broader loss: we've dropped the tradition of thinking while moving — commuting seated, working seated, consuming seated. Walking doesn't need to be justified as a productivity tool; the body in motion is itself how thoughts get re-sorted. Go outside today, no earbuds, no podcast, and let your feet organize your mind.
工具角Tool Corner
- Glashütte 垃圾钟表:只走 30 分钟的机械钟Glashütte Trash Clock: A Mechanical Clock That Runs for 30 Minutes
- 第一步:艺术家 Niklas Roy 在驻留德国钟表重镇 Glashütte 期间,用当地捡来的垃圾和杂件做了一台完全可运行的机械钟——站在顶级制表业的故乡,用废料向它致敬。Step 1: While in residence in Glashütte — Germany's watchmaking capital — artist Niklas Roy built a fully working mechanical clock entirely out of trash and found objects. Standing in the hometown of precision horology, he pays tribute to it with garbage.
- 第二步:机芯故意「不精确」:整机只运行约半小时,只显示分钟和秒,分针每过十二点点一次锣,而且能量耗尽前会触发一个外部「随机发生器」——一台知道自己会停摆的钟。Roy 还顺手发明了「GTC 时间标度」,让这台垃圾钟充当「全球参考时钟」,调侃当地对「Glashütte/SA」产地标识的严格法律。Step 2: The movement is proudly imprecise: it runs for only about half an hour, displays just minutes and seconds, strikes a gong whenever the minute hand passes twelve, and triggers an external randomizer just before it dies down — a clock that knows it will stop. Roy also invented the 'GTC time scale,' making the trash clock a global reference clock, a wink at the strict local law protecting the 'Glashütte/SA' hallmark.
- 第三步:家史是暗线:Roy 的曾祖父和曾曾祖父都是瑞士钟表匠,从 La Chaux-de-Fonds 迁到萨克森继续制表。用垃圾在 Glashütte 造钟,是钟表匠后人对精密传统的一次倒置式传承——值得看的还有整个机构逐件拆解的文档,全手工齿轮链的代码级可读性。Step 3: The hidden thread is family: Roy's great-grandfather and great-great-grandfather were Swiss watchmakers who moved from La Chaux-de-Fonds to Saxony. Building a Glashütte clock from trash is a horological heirloom inverted — and the build documentation, piece by gear by piece, reads like clean code.
- FlickBoard:拒绝帮你改字的手机键盘FlickBoard: The Phone Keyboard That Refuses to Fix Your Words
- 第一步:作者的抱怨具体到让人共鸣:手机屏幕装不下完整的 QWERTY,于是键盘厂商塞进了预测和自动更正作为「补偿」。但他只想要「按下键,屏幕上就出那个字母,该死」——每一次被自动改字,都是工具替你做了一次你没同意的决定。Step 1: The author's grievance is painfully concrete: phone screens can't fit a full QWERTY, so keyboard vendors stuff in prediction and autocorrect as compensation. All he wants is 'I press a key, the letter shows up, goddammit.' Every auto-fix is the tool making one decision you never agreed to.
- 第二步:解法是把 12 个键压成单排 QWERTY:布局保持熟悉,符号与数字用滑动(flick)手势共享同一排键。不是更聪明的键盘,而是更诚实的键盘——它放弃预测,换来你对每一次击键的完全控制。Step 2: The answer: a twelve-key single-row QWERTY. The layout stays familiar; symbols and digits share the row via flick gestures. Not a smarter keyboard — a more honest one: it gives up prediction in exchange for your complete control of every keystroke.
- 第三步:读这篇文章的价值大于用这个键盘的价值:它示范了一种对抗「AI 默认接管」的设计立场——先承认「我就是一个字一个字的人」,再为此造工具。博客本身就是这个立场的作品:带脚注、带自嘲、不讨好任何算法。Step 3: Reading the essay matters more than installing the keyboard: it models a design stance against AI-default-takeover — first admit 'I am a person who types one letter at a time,' then build for that person. The blog post itself is the stance's artifact: footnotes, self-mockery, and zero pandering to any algorithm.
- Lathoa:一个故意算错的 AI 数学老师Lathoa: The AI Math Tutor That Gets It Wrong on Purpose
- 第一步:产品做了一次干净的反转:所有学习 App 都在教孩子「从 AI 那里拿到正确答案」,Lathoa 的机器人 Errol 却每一步都大声演算、并故意在其中一步犯错——解「电影 3:45 开始放映 90 分钟」时,它把 90 分钟算成 0.90 小时,得出 4:35 的荒谬答案。Step 1: The product executes a clean inversion. Every learning app teaches kids to extract the right answer from AI; Lathoa's robot Errol works each problem out loud and gets exactly one step wrong on purpose — for 'a movie starts at 3:45 and runs 90 minutes,' it converts 90 minutes into 0.90 hours and lands on the absurd 4:35.
- 第二步:孩子的角色从抄答案的变成破案的:读步骤像读案卷、点出那一步出错(机器人「自信度」随之肉眼可见地垮掉)、再用自己的话解释错在哪里,才升侦探等级。错误的迷惑度分三档(明显 / 狡猾 / 专家)按年龄校准——出题者变成了「会犯错的人」,而这正是真实世界里的同事、记者和老板。Step 2: The kid's role flips from answer-copier to detective: read the steps like a case file, tap the broken one (the robot's on-screen 'confidence' visibly deflates), then explain in your own words why it breaks to earn rank. Errors come in three calibrated tiers — obvious, sneaky, expert — turning the question-setter into a fallible human, which is exactly what colleagues, reporters, and bosses are.
- 第三步:点子的原创之处在于对 AI 时代的诊断:孩子将来每天听到一百个「自信的错误答案」,防御它靠的不是记住更多正确答案,而是习惯审视论证本身。把「抓 AI 的错」变成奖励机制,等于给不信任装了入门教程——这可能是今天最有未来感的一个产品想法。Step 3: The originality sits in its diagnosis of the AI era: kids will hear a hundred plausible-sounding wrong answers a day, and the defense against them isn't memorizing more right answers — it's the habit of auditing arguments themselves. Rewarding kids for catching an AI's mistake is an onboarding tutorial for distrust. Easily the most future-aware product idea in today's corner.